CT-based deep learning model for the prediction of DNA mismatch repair deficient colorectal cancer: a diagnostic study

Colorectal cancer (CRC) is the third most commonly diagnosed malignancy in the world and the second highest rate of increasing incidence among all gastrointestinal tumors [1, 2]. Patients with DNA mismatch repair deficient (dMMR)/microsatellite instability-high (MSI-H) CRC have a short overall survival (OS) and obtain no benefit from adjuvant chemotherapy [3‐5]. More importantly, recent studies have demonstrated dMMR/MSI-H is a predictive biomarker for immunotherapy, because dMMR/MSI-H CRCs are associated with a higher mutational burden tumor neoantigen load, and dense immune cell infiltration [6, 7]. In addition, immunotherapy in patients with advanced CRC harboring dMMR/MSI has been approved by the United State Food and Drug Administration (FDA).

The National Comprehensive Cancer Network (NCCN) and the European Society for Medical Oncology (ESMO) guidelines both recommended that all patients with CRC be tested for microsatellite instability, a hypermutable phenotype caused by defects in DNA mismatch repair, which facilitated individualized clinical making-decisions, then maximized the benefits for CRC patients [8‐10]. Current testing for dMMR/MSI include PCR-based assay for microsatellite markers and immunohistochemical analysis for MMR protein expression [11]. While these existed approaches face distinct drawbacks. First, routine MSI testing using IHC or PCR is not commonly performed on account of tedious procedures and the heavy financial burden [12, 13]. In addition, the procedure of sampling is invasive linked to potentially complications, which limits the dynamic monitoring of biological characteristics and histopathological changes of tumors [14]. Furthermore, the accuracy of conventional biopsy specimens will be influenced by sampling errors, such as insufficient or inappropriate tissue sampling because of tumor heterogeneity. Therefore, a noninvasive and accurate method is highly desirable for pretherapeutic prediction of MMR status, to help better stratify CRC patients before individualized clinical making-decision.

Recently, artificial intelligence (AI) algorithms, particularly deep learning, have shown outstanding performance in medical image processing, advancing the field forward at a rapid pace [15]. A typical approach of deep learning termed convolutional neural network (CNN) has been reported to act as an alternative tool to tackle complex medical issues efficiently and effectively and achieve satisfying performance in many diseases such as breast cancer, pulmonary nodules, gastrointestinal cancer and CRC [16‐19]. Recent studies have proposed deep learning approaches for predicting MMR status based on hematoxylin and eosin histological images in CRC patients [20‐22]. These studies reported a moderate predictive performance, which suggested that it is reasonable to speculate that deep learning approaches can achieve improved performance for pretherapeutic prediction of MMR status in CRC. However, few studies based on deep learning focus on routine and noninvasive computed tomography (CT) images.

In the present study, we aim to develop and validate a deep learning model based on pretherapeutic CT images for predicting MMR status in CRC. We hypothesized that DL model could make it easier and more precisive to stratify MMR status, which could promote the personalized clinical-making decision.

We developed a fully automated classifier for stratifying MMR status in 1812 patients with CRC using clinically acquired pretherapeutic CT images from two institutes. It demonstrated promising performance with an AUC of 0.986 in the internal validation cohort; moreover, when further validated in an external validation cohort, it demonstrated robust performance, with an AUC as high as 0.915. It’s noted that the MMRnet based on fully automatic deep learning successfully triaged MMR status in different groups by subgroups analysis. The outperformance of the MMRnet model indicated that CT-based full-automatic deep learning could serve as a noninvasive tool for the pretreatment prediction of MMR status in CRC, further enabling the clinical implementation of computer-aided personalized management for CRC patients.

An architecture of networks was designed by paralleling a max-pooling layer with a center-cropping layer to extract information from different scales. In our study, the Resnet was used for image processing, one of the most popular deep learning models for image analysis, which ease the training of networks that are deeper than those used previously [25]. In addition, we applied the GPR model to fuse the information of different axial images for automatic interpretation results of CT images, which can reduce computational complexity and improve the accuracy of fusion results to a certain extent compared with the general linear model. Our results displayed excellent performance using the DL model to stratify MMR status. Although the underlying mechanism of using deep learning to predict MMR status remains unclear, we hypothesized that this could be related to tumor heterogeneity. After all, the widespread application of deep learning in the non-invasive analysis of tumor heterogeneity in the field of oncology has been demonstrated in many previous studies [26, 27]. Additionally, it is reported that dMMR/MSI-H tumors tend to have distinct morphological patterns such as poor differentiation, mucinous differentiation, histological heterogeneity, infiltrating lymphocytes, and significant Crohn-like reactions at the tumor frontier [28, 29]. These histopathological features imply that the dMMR tumors may be more heterogeneous than pMMR tumors, which could be captured by the DL model. In our study, the DL model achieved excellent performance in MMR status prediction of CRC patients, which was comparable to and even superior to the results of previous studies. It was indicated that deep learning could gain more high-dimensional image information about tumor heterogeneity that cannot be captured by human eyes [20‐22]. In addition, CT-based radiomics are predominant in image-based models and have been proposed to preoperative discriminate dMMR and pMMR in colorectal cancer [30‐32]. However, tumors were usually segmented manually in majority radiomics researches, which was time-consuming and inevitably caused inter-observer variations. These defects could be avoided by the DL method, and it could adaptively extract features according to the data rather than using predefined features.

In the statistical analysis, we found the age and tumor location present statistical significance in the differentiated dMMR group and pMMR group, indicating that the younger patients and the tumor located in the right colon were more likely to show dMMR status, which was consistent with previous studies [33]. Our study identified that dMMR CRC occurs predominantly in the younger population and on the right side, which may be explained by the fact that the proximal and distal colon have different embryonic origins, leading to distinct biological properties [34, 35]. In the present study, the clinical parameters were not incorporated in the final prediction model, mainly based on the following considerations. Firstly, the fully automatic DL model had outperformed discriminative performance in stratifying MMR status, with an AUC of 0.986 in the internal validation cohort. Meanwhile, the predictive performance was robust in the external validation cohort, which achieved an AUC of 0.915. Hence, we have reason to believe that the DL model developed in the present study could be considered an independent tool to predict the MMR status for CRC patients. In addition, due to the excellent predictive performance of this DL model, the actual predictive efficacy of clinical parameters may be obscured when incorporated into the DL model. Therefore, we tend to construct a relatively simple and feasible DL model rather than a combined model incorporating the predictive clinical parameters. For the clinical parameters, subgroups analyses were performed in addition to the main analysis to assess the prediction performance in different subgroups in this study, and the result revealed that the MMRnet model showed similar satisfying prediction performance in all groups, that indicating the pre-therapeutic CT-based DL model could potentially serve as an alternative approach to predict the MMR status, then help in clinical decision-making for CRC patients.

Our DL model has unique advantages. First, the imaging data could be stored and used repeatedly, in addition, the noninvasive deep learning method is more convenient than the pathological approach based on biological specimens. Secondly, we should be cognizant that deep learning models can produce study results more directly and quickly and improve the problem of being time-consuming and burdensome for busy clinicians compared with ROI-based analysis. This may be attributed to the automatic feature extraction based on deep learning models probably have the ability to capture additional differences within tumors [36, 37]. Moreover, we enrolled a higher sample size of over 1800 CRC patients, which met the requirement of millions of weights to train efficiently for CNN, and our model demonstrated superior performance to previous radiological models in terms of AUC values (0.74–0.82 for other models) [31, 38, 39]. In conclusion, it is suggested that the DL model based on CT images has the potential to stratify tumor MMR status with promising effectiveness prior to surgery, chemotherapy, or immunotherapy in our study.

Despite promising findings, our study has some limitations. First, the retrospective nature of this study inevitably leads to selection bias, and a prospective and multicenter study is required to confirm the impact of our model in the future so that it could serve better in clinical application. Second, although we had a favorable predictive performance in participants from an external institution, it did not reach a high level in the internal institution. We considered that no specific schemes were applied to deal with the parameter variations from different scanners. Third, we explored the DL model based on routine CT images rather than other imaging technologies, such as MRI and Dual-Energy computed tomography, which have been reported in previous studies [39, 40]. Actually, CT is the most suitable and routine examination method for colorectal cancer. Finally, since there is no specific definition of these extracted features of deep learning, their interpretability should further explore image encoding processes in the future.

We have constructed and validated the fully automatic DL model derived from pre-therapeutic CT images, which can stratify MMR status in CRC, with superior performance. This method could provide a potential noninvasive tool to triage MMR status in CRC, thus further personalized medicine.